This is my first attempt at a YouTube video, and I got it done, at least. This was material to try to help my friends out that are using AI and trying to get better with it. The gist of the video is to keep all your output from AI and your input for AI in the same folder. That way, your system builds on itself and gets smarter, and AI is able to help you with your particular needs for any of the contents within the folder. The folder may grow in scope and size over time, as mine has, and then I split it out into a separate knowledge base for information that doesn’t pertain to anything else in the initial folder.
Having Claude Co-work use a folder as a project
I currently I have a few folders set up as projects, and they ended up becoming AI knowledge bases. They started out as a means of organization to track receipts and invoices, configurations, and settings. There was some reorganization, and the organization continues to evolve over time with usage and growth of information stored in my folders. And of course, with organizing, you keep different subject matters from being placed in the subfolder for a completely different subject matter.
I kind of split my folders as projects into separate subject matters. Examples are:
- Work documents, issues, trackers, configurations
- Business documents for rentals
- Stock information and trades and research, as well as stock trading methodologies from traders such as Kristjan Qullamaggie and Mark Minervini
This is not an exhaustive list of my different projects, but a small subset, just as an example. When it is time to get to work on these, I can open up the folder in Anthropic’s Claude co-work. I can use Claude in the console, so Claude Code, or switch to VS Code to use Claude on my folders.
Switching projects from Claude.ai to Codex (ChatGPT)
Another benefit of keeping all your context inside of a folder to work with in skills as you build them is you can switch which AI systems you want to use. For example, I mostly work in Claude, but if I wanted to, I can open the same folder in Codex. Codex should be able to use the Claude skills and Claude Markdown instructions in a claude.md if I initialize the folder to know exactly what’s in the project in the full context of the system. You can even jump into VS Code or command-line Claude if you really want to. Claude Code.
As I’ve been testing this since about March, Claude can use Codex’s .agent.md files, and Codex can use claude.md files as well.Codex now knows about skills, so it can use Claude skills to pick right up where we left off when I was in my Claude session, if I’m now working in Codex desktop.
My friends would tell me how excited they were that they were playing with AI now, and they wanted to know how to get better at it. The first advice I give them is to use a desktop app, share a folder, and keep all your knowledge and all your work in that folder so that AI has access to it.
Ask a general-purpose AI chatbot to answer a customer email and you will usually get something that
reads beautifully and is partly false. In this video I paste a real-sounding email from a customer whose
air conditioner is blowing warm air, and the AI writes back a polite, organized reply that claims we
installed her system three years ago. We had never met her. There is no such customer, and there is no
such company. The AI simply wrote what a heating and air company would probably say, and said it as us.
What makes that dangerous is not the invented sentence. It is that the same reply also quoted an $89
diagnostic fee, which is roughly the real going rate. You cannot look at the answer and tell which half
came from knowledge and which half came from guessing.
The fix takes under ten minutes and costs nothing. Work out of a folder. Get a desktop AI tool, point it
at one folder on your computer, and put what your business actually knows into plain text files inside
it: your services, your pricing, your frequently asked questions, your procedures. I build exactly that
for a fictional heating and air company using three files, then ask the same customer question again.
This time the reply uses our real time frames, our real emergency dispatch policy, and our real rates.
The folder also becomes your history. You stop re-explaining who you are at the start of every chat, and
the next session picks up where the last one left off.
The problem: confident invention
- A generic chatbot given a customer email produces a reply that is well organized, polite, and sounds
like it came from a real company. The tone is never the giveaway. - In the demo it asserts “because we installed your system 3 years ago.” That never happened. The AI has
never heard of the customer (Rhonda) or of the business. - It writes what a heating and air company would plausibly say, then says it in first person, as you.
- It also quotes an $89 diagnostic fee, which is close to the real average. That is the actual hazard:
correct and invented details arrive in the same paragraph, in the same confident voice, with nothing
marking the difference.
The fix: use a folder as the project
What “working out of a folder” means
- Install a desktop AI tool (Phil names Claude or Codex) and point it at one folder on your machine.
- Everything you do lives in that folder: the source material, the notes, and the results.
- You control what the AI can see on your system. It reads the folder, not your whole computer.
- Over time the folder stops being scratch space and becomes a knowledge base you can question.
Why this replaces prompt engineering
- The old approach was to paste all the background into the chat first, then ask the question. That was
prompt engineering, and you repeated it every session. - With a folder, the context lives on disk. The AI looks through the folder, finds what it needs, and
answers. You do not re-explain the project every time. - Prompt-style front-loading still matters where there is no folder or harness behind the model, for
example AI voice agents.
One folder per subject
- Keep unrelated work in separate folders so the AI does not blend subjects together.
- Phil runs about six: a work repository with a configuration database, daily notes, issues and project
tracking; receipts and purchases for reconciliation; a rental property; stocks; books and research. - He also keeps one deliberate hodgepodge folder for things that are not developed enough to deserve
their own project. When a topic graduates, it moves out into its own folder.
The demo: Mike’s Heating Company
- The fictional company folder holds three files covering services, pricing, and frequently asked
questions (FAQs). That is the whole knowledge base. - The test input is a customer email: Rhonda Kay writes that her air conditioner has been blowing warm
air since yesterday, it is 104 degrees out, and how soon can someone come out. - Pointed at the folder, the AI drafts a reply that uses the company’s actual response time frames, its
actual rates, its emergency dispatch policy, and the emergency surcharge, because the surcharge is a
documented policy in the folder. - The answer is grounded because the facts were gathered once and written down, not because the model
got smarter.
What the folder still does not know
- The warranty line (“should be 10 years”) is a generic assumption. There is no customer relationship
management (CRM) system connected, so the AI cannot confirm whether this customer is actually covered. - Phone numbers are missing from the reply because nobody has put them in the notes yet.
- The lesson is the useful half: the gaps are now visible and fixable. Anything absent from the folder is
something you can go add, instead of something the model quietly invents.
Why markdown, and why Obsidian shows up
- All three files are plain text, specifically markdown.
- Markdown is a simple text format, not binary like a Word document, so both the AI and your computer can
read it directly. - The markup characters are just formatting hints (a hash mark makes a header, for example), the same way
HTML source looks ugly but renders cleanly in a browser. - Obsidian appears in the video only as a front-end viewer that renders those markdown files in a
readable way. It is not required, and Phil explicitly calls it “not important at all.” - The practical rule: markdown and text files are the easiest things for AI to read right now.
Getting your own folder started
- Gather what the business knows and write it down: procedures, documentation, services, pricing, FAQs.
- If a standard operating procedure (SOP) only exists in one employee’s head, that is the first thing to
write into a file. The system cannot use context it was never given. - The same method works outside business. For household tracking, record purchase dates and serial
numbers for appliances such as the washing machine and refrigerator, so a warranty call has something
to draw on. - The folder is never finished. Each session adds a little more.
Key takeaways
- A generic AI answer will mix real facts and invented ones in the same confident paragraph, and you
cannot tell them apart by reading it. - The cheapest fix is a folder: plain text files holding what you actually know, with an AI tool pointed
at it. - Context on disk beats context retyped into chat. You stop re-explaining yourself every session.
- Keep unrelated projects in separate folders so subjects do not bleed into each other.
- Three files were enough to change the quality of the answer. This is not a big-project technique.
- Gaps in the folder become visible gaps in the answer (no CRM, no phone numbers), which is far better
than invisible invention. - Markdown is the preferred format because it is plain text; a viewer such as Obsidian is optional polish.
- Next video in the series: point a different tool (Codex) at the same folder and have it pick up the
project with no re-explanation.